List of Papers Browse by Subject Areas Author List
Abstract
Accurate multi-class segmentation of the aorta in medical CT images is essential for the effective diagnosis and treatment of blood flow abnormalities.
However, achieving precise segmentation in multi-zone remains challenging due to the lack of visible boundaries and the similarity in intensity between zones.
Although existing methods incorporate anatomical features such as global geometric constraints and landmark-based alignment, they often struggle when these features are difficult to extract, such as in regions with asymmetric deformation or extreme curvature due to dissection.
This limitation of relying solely on simple anatomical cues underscores the need to learn and model complex anatomical interrelationships for robust segmentation.
To overcome these challenges, we propose a plane detection-based segmentation framework that is constrained by anatomical features and their relationships to accurately detect planes between zones.
Specifically, our method detects planes by localizing centerpoints and regressing the corresponding normal vectors, while anatomical landmarks further refine the position and orientation of these planes. Additionally, anatomical regularization losses enforce geometric consistency among these components, thereby enhancing both accuracy and stability of the detected planes.
The entire framework is implemented as an end-to-end architecture, enabling efficient learning.
The experimental results on the AortaSeg24 dataset demonstrate that our approach achieves state-of-the-art performance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3576_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jjong0225/ACP
Link to the Dataset(s)
AortaSeg24: https://aortaseg24.grand-challenge.org/
BibTex
@InProceedings{AnJon_Aorta_MICCAI2025,
author = { An, Jonghoon and Lee, Dong Hyun and Kim, So Hyun and Moon, Taejin and Chung, Minyoung},
title = { { Aorta Multi-class Segmentation via Anatomically Constrained Plane Detection } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15962},
month = {September},
page = {56 -- 66}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a method for multi-class segmentation of the aorta in medical CT images which is relevant for the diagnosis and treatment of blood flow abnormalities. To tackle this challenge the authors propose a plane detection-based segmentation framework that is constrained by anatomical features and their relationships. The work is carried out on a publicly available cohort (AortaSeg24) which allows to put the results into perspective. Achieved results are claimed to be superior compared to state of the art.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- In general, the end-to-end approach is interesting. While it increases overall complexity to tackle the task, the fact that anatomical features and their relation is part of the pipeline may increase robustness in scenarios of uncertainty and the overall optimal solution may be found. Of course, when combining everything into one pipeline, there might be the potential that it will be difficult to interpret the prediction result which may be clinically relevant. However, I anticipate that the individual anatomical features can be extracted to verify the prediction result.
- Clarity: The paper is well structured and can be followed accordingly. Proper use of figures made that help understanding the paper. The contributions are properly highlighted.
- Use of public cohort to allow comparison and benchmarking of results.
- Ablation study carried out to assess the benefit of the individual contributions.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- Data cohort: Though it is certainly a strength that a public cohort was used for development, the fact that this is the only data used also limits the overall impact. Further, evaluation is required to understand potential limitations of the method and show generalization of the results.
- Limited reproducibility: Unfortunately, the authors do not state to make their code public pending approval of the paper. Though this is certainly not a must, the limitation is that for such end-to-end pipeline there are significant limitations if others want to reproduce results. The architecture is quite complex with many details not covered in the paper due to space limitations. At the end, the result is that others may have difficult time to re-implement the paper and reproduce results. This also means that the approach can not be used by others as benchmark. In particular, I find this a pity as components may also be interesting for other similar problems.
- Complex loss: Putting everything together results in quite an extensive loss with several weighting factors to balance the individual contribution. This immediately raises the question how sensitive the approach to the weighting, how parameters were found etc. At the same time, the ablation experiment is appreciated.
- Evaluation: Personally, I am very much struggling with the level of evaluation carried out though the paper follows typically tendency. First of all, I find it a drastic simplification to limit the evaluation to a simple value (mean) and not indicate further measures (e.g., std, min/max). It is essential to now if the method in general provided better result or if maybe 1-2 drastic outliers got improved. This is simple additional information, does not require a lot of space, but adds a lot of information. Results are provided with 3-4 digits accuracy. Is this a reasonable granularity? Is this level required? This brings me to the last point - at this level, I really think that statistical significant tests are needed. Though it is appreciated to report the increase in performance and also it relative increase in percent, it is important to perform statistical testing. From my perspective this should become more state of the art in evaluation given the accuracy levels achieved.
- Comparison to state of the art: Doing a comparison to general approaches as done in the paper with comparison to UNET and SWINUNETR is always to be taken with caution. It triggers the question how the implementation was done, how were parameters chosen etc? Was it optimized properly? nn-Unet shows that it is less about the architecture, but a lot how it is adapated. Why is there no benchmark against nn-Unet?
- Pipeline versus individual approach. The proposed idea is to learn all the individual components that define the regions in an end-to-end fashion. At the same time, one may also think to tackle the problem in a hierarchical/modular fashion based on the components that define the zone (e.g., independent landmark detection, centerpoint detection etc.) From my perspective, it would be interesting to benchmark against such more modular approach to appreciate the true value of a combined approach.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
- The introduction is well written - though potentially a bit too extensive (maybe you leave the summary of contributions at the end as it is a bit repetitive). You may want to consider shortening to leave more space for other relevant sections. The figure illustrating the overall approach was very helpful and allowed to quickly understand the problem and the approach. Motivation is clearly stated including short clinical background. Related work is cited adequately. Not sure what A5, A2 were in the introduction and why it is put like this in the text (is it anonymization)? Please change in the final version.
- You claim that other approaches have shortcomings when “features are poorly extracted such as in regions with asymmetric deformation or extreme curvature due to dissection”. However, I wonder if in such pathological scenarios the anatomical ground truth is also well defined. How does your approach overcome this challenge and please more accurate highlight improved performance on such cases (e.g., tailored evaluation).
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
In general, a solid paper with an interesting approach. At the same time, the paper also comes with several weaknesses that limit its impact which are in summary: 1) Limited reproducibility - it will be difficult for others to reimplement and as such only the achieved performance results are to be cited potentially. 2) Limited cohort - only focus on AortaSeg so it remains unclear how achieved results generalize 3) Insufficient rigor in evaluation - from my perspective, there are several flaws in the evaluation which limit confidence in the overall superiority of the approach.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Too my mind, a very constructive rebuttal by the authors. To ensure reproducibility, they authors now committed themselves to publicly release the source code upon final acceptance. Furthermore, the evaluation has been extended to also include std and min-max to confirm that the improvements are not only driven by outlier handling, but show a consistent improvement.
I would highly encourage the authors to evaluate the method on further independent cohorts and show generalization of the results.
Review #2
- Please describe the contribution of the paper
This paper presents a plane detection-based segmentation framework that is constrained by anatomical features and their relationships to accurately detect planes between aortic zones
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper is well written, clearly structured, and presents a well-motivated clinical and technical problem. The authors effectively communicate their contributions and the relevance of their work within the medical imaging field
- The proposed methodology appears technically sound and demonstrates clear advantages compared to the current state-of-the-art.
- In addition to the segmentation task, the authors provide a clinical assessment by measuring anatomical diameters, which reinforces the potential practical impact and applicability of the method in a clinical setting.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
There are issues regarding the clarity of the methodology that should be addressed to improve the overall understanding and reproducibility of the work:
- It is not clearly stated what the output of the segmentation module is. Does the network produce a full segmentation of the entire aortic structure, or does it segment individual anatomical regions separately?
- During training, it is unclear whether the predicted planes are used to partition an existing full segmentation mask based on plane detection.
- The implementation of the anatomical regularization term also lacks clarity. Is the regularization applied to ensure that each individual plane is anatomically consistent, or does it enforce consistency across all detected planes collectively? This distinction is important and should be clarified in the manuscript.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
- To ensure the reproducibility of the proposed method, the authors should clarify the methodological details highlighted in the weaknesses section.
- It would be helpful for the authors to clarify whether the network modules are initialized from scratch or if any pre-trained models are used to ensure stable convergence and improved performance.
- As a suggestion for future work, the authors could discuss incorporating a consistency loss between the multi-label segmentation output and the detected planes.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a relevant clinical problem with a well-structured and technically sound approach. The results are good. However, there are several methodological aspects that lack clarity. If the authors are able to clarify these points and provide additional methodological details, the paper could be a valuable addition to the field.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The purpose of this paper is to develop a framework for the automatic segmentation of the aorta. This work includes not only the segmentation of the aorta as a whole but also the delimitation of its different segments after the segmentation process. To delineate the nine proposed segments, seven landmarks corresponding to the root of the branches of the aorta and their corresponding points on the aortic centerline are used. These landmarks and the corresponding centerline points are automatically detected using convolutional neural networks (CNNs). From these points, the optimal planes separating the different segments are obtained. These optimal planes are those whose normal vector is 90 degrees relative to the line joining the landmark with the center point. The method has been tested using the AortaSeg24 Challenge Database. For evaluation, the Dice Similarity Coefficient (DSC) percentile, Hausdorff distance (HD95), and average symmetric surface distance (ASSD) are used to measure different aspects of the segmentation process. Finally, the method has been tested against various state-of-the-art (SOTA) algorithms, yielding excellent results. The main contribution of this paper is the proposal of a new aortic segmentation framework based on plane detection using anatomical features.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The developed method demonstrates significant robustness, as evidenced by the results obtained from testing on a generic database and subsequent comparison with state-of-the-art (SOTA) algorithms. The paper includes an ablation study that helps readers understand the impact of the different components of the proposed method
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The method relies on the accurate detection of anatomical landmarks related to the branches of the aorta. However, these landmarks are not exactly the standard ones referenced in clinical guides used by specialists (like reference 4 of the current paper). Nevertheless, the method could be easily adapted to these more standard segments used by specialists. Another issue is the optimization based solely on a rotation angle. Since this involves a 3D plane, two angles should be used for such optimization.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This is a very comprehensive work with excellent results. These results have been tested on a generic database (AortaSeg24) and compared with SOTA algorithms, making it an outstanding contribution
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed many of the concerns raised by the reviewers. However, the issue of reproducibility remains unresolved. Although the results were obtained using a generic database (AortaSeg24), the authors do not appear to intend to release the code, even if the work is accepted. This is undoubtedly a limitation.
Author Feedback
General Response: We sincerely thank the reviewers for their constructive feedback and thoughtful critiques. In this rebuttal, we clarify key aspects related to implementation details, evaluation fairness, and the dataset used. We believe several concerns may stem from misinterpretations, and we aim to resolve them through the following responses. To ensure reproducibility, we will publicly release the source code upon final acceptance.
Data Cohort and Reproducibility: We used the AortaSeg24 dataset, the only publicly available resource for multi-zone aortic segmentation. Although it constitutes a single cohort, it was acquired across multiple local imaging centers using varied scanners and imaging protocols, and annotated by clinical experts [6]. While we acknowledge the limitation of using a single dataset, acquiring multi-centric manual annotations was not feasible. We believe AortaSeg24 offers a reasonable and fair benchmark for academic evaluation.
Our framework was designed as an end-to-end architecture to jointly optimize complementary multi-task objectives. While this introduces complexity in loss formulations and potential training instability, it is necessary to achieve the reported accuracy. As with most recent models, hyperparameter tuning is dataset-dependent and often performed manually for each method; current automation strategies such as meta-learning are still limited in practical adoption. Due to space limitations, ablation studies of the end-to-end pipeline and loss could not be included in the paper but will be provided with the released code.
- Evaluation Protocol: Our comparative evaluation includes nnU-Net framework, a strong state-of-the-art baseline that leverages automated adaptation and tuning. Specifically, we evaluated A5, an nnU-Net variant augmented with skeleton recall loss tailored for aortic segmentation. All other competing models were trained under identical conditions as our method and rigorously tuned, ensuring fair comparisons with their best configurations.
As shown in Table 1, the reported mean metrics clearly demonstrate the superior performance of our method. To address concerns regarding statistical robustness, we provide additional summary statistics for Table 1, including the standard deviation, minimum, and maximum values of DSC, HD95, ASSD, and DE. Due to space constraints, statistical analyses were limited to Baseline, SOTA, and Our model. The complete results will be provided in the camera-ready version.
Method DSC HD95 ASSD DE Baseline (SwinUNETR) 0.72 ± 0.057 (0.63-0.80) 9.04 ± 3.00 (5.02-13.85) 2.48 ± 0.76 (1.52-3.89) 5.47 ± 2.03 (2.89-9.85) SOTA (A5, nnU-Net variant model) 0.75 ± 0.045 (0.69-0.81) 7.59 ± 2.45 (4.77-12.29) 2.16 ± 0.57 (1.50-3.23) 4.00 ± 1.18 (2.62-5.80) Ours 0.79 ± 0.039 (0.73-0.85) 5.99 ± 2.06 (3.50-9.59) 1.75 ± 0.43 (1.19-2.65) 3.02 ± 1.11 (1.64-5.31) These results confirm that our improvements are not driven by outlier handling, but reflect consistent performance gains across cases, including asymmetric deformation or extreme curvature.
- Implementation Clarifications: Our single segmentation module produces 14 multi-class segmentation logits, covering one unified zone class and 13 artery-specific classes (section 2.4). Anatomical regularization is applied independently per plane, as noted in Eq. 3. Please refer to Fig. 2(c) for a visual illustration of how the regularization is applied to each plane. Importantly, this plane-based separation is used only during inference, not during training. All models were trained from scratch using Xavier initialization, without any pre-training. The full implementation details, including the pipeline settings, will be made clear through the released code.
We appreciate the reviewers’ comments and will incorporate all necessary revisions and reference updates in the final camera-ready version.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
The work proposes a segmentation method for multiple aorta subsections based on predicting delimiting planes between segments. The approach is clinically relevant and an unsolved problem.
Reviews are mixed and a number of important weaknesses are mentioned. Authors are invited to organize these weaknesses appropriately and respond in a rebuttal.
From my point of view, especially the concerns about data cohort size, reproducibility due to limited implementation details, limitations in evaluation and reported measures, as well as comparison with state of the art must be addressed.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper presents a well-motivated and technically sound method for multi-class aortic segmentation via anatomically constrained plane detection. All reviewers support acceptance, noting the strong performance, clinical relevance, and thoughtful design of the framework. While reproducibility was initially a concern, the authors have now committed to releasing the full source code upon acceptance, which addresses the main outstanding issue. The evaluation is conducted rigorously on the AortaSeg24 dataset using strong baselines, including a tailored nnU-Net variant. Additional metrics and detailed clarifications strengthen the validity of the reported improvements.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A